logoJadeHouse Lab
Background Image

lwrLimbsEMG

Lower Limb Analysis Based on Surface Electromyography (sEMG) Using Different Time-frequency Representation Techniques

17 March 2024

Citation

Mai, Ramadan, Ibraheem. (2023). Lower Limb Analysis Based on Surface Electromyography (sEMG) Using Different Time-frequency Representation Techniques. International Journal on Advanced Science, Engineering and Information Technology, 13(1):24-24. doi: 10.18517/ijaseit.13.1.16685 [Link to paper]

Summary

Research Question

  • Investigate the effectiveness of different time-frequency representation techniques (scalogram, spectrogram, persistence spectrum) in mapping 1D sEMG signals from lower limb muscles onto 2D images for the detection of knee abnormalities.

Research Conducted

  • Comparative experiments using three time-frequency representation techniques (scalogram, spectrogram, persistence spectrum) to analyze lower limb muscle activity using surface electromyography (sEMG) signals.

Key Results

  • Demonstrated that the scalogram technique outperformed spectrogram and persistence spectrum in recognizing knee abnormalities.
  • Convolutional neural network (CNN) model used for detecting knee abnormalities showed improved performance when fed with 2D projected images from sEMG signals.

Significance

  • Proposed technique could serve as a clinical tool for diagnosing various muscle activities, potentially enhancing the accuracy of such diagnoses.
  • Developed 2D deep CNN model can achieve more accurate classification than in classical learning approaches via manually extracted features.

Methods

Data Collection

Acquisition

Four electrodes were placed on specific muscles and a goniometer on the knee to produce time series data for each muscle during various movements. The muscles analyzed include:

  • Rectus Femoris (RF): A muscle in the quadriceps group located in the front of the thigh.
  • Biceps Femoris (BF): Part of the hamstring muscles, located at the back of the thigh.
  • Vastus Medialis (VM): Also part of the quadriceps group, situated medially in the thigh.
  • Semitendinosus (ST): Another muscle in the hamstrings group, positioned at the back of the thigh.

Sample Size and Characteristics

22 samples were used, with 11 normal and 11 exhibiting knee pathology. Three different movements per subject resulted in 66 records.

Data Preparation

Signal Preparation and Pre-Processing

  • Data for each muscle attribute was gathered into 10 muscle data files for both normal and abnormal measurements.
  • Preprocessing was performed to remove any artifacts while minimizing signal loss.

Data Augmentation

  • A time-series generator was used to extend the dataset, resulting in 1056 records of sEMG segment signal for the corresponding muscles.

2-D Mapping Generation

1-D sEMG signals were converted into a 2-D time-frequency space using three techniques:

  • Scalogram: Utilizes Continuous Wavelet Transform (CWT) to analyze sEMG signals in both time and frequency domains, providing a multi-scale analysis.
  • Spectrogram: Employs Short-Time Fourier Transform (STFT) to map 1-D signals into 2-D representations, offering spectral information at various time segments.
  • Persistence Spectrum: Based on the time percentage a given frequency persists within the signal, useful for tracking and extracting signal ridges in time-frequency maps.

Lower Limb Abnormality Characterization

Convolutional Neural Network (CNN) Development

  • A 12-layer CNN model was utilized to perform binary classification (normal/abnormal) using the 2D structure of the mapped input images.
  • Four convolutional layers were followed by corresponding pooling layers, leading to three fully connected layers for accurate learning and deep feature extraction.
  • Hyperparameter tuning was conducted based on loss values to optimize performance and prevent overfitting, with parameters such as the number of kernels, number of fully connected layers, dropout parameters, and ReLU activation function.

Model Training

10-fold cross-validation was used to assess the model’s performance during hyperparameter tuning.

Results

Scalogram image representation of sEMG signals provided significantly better performance in recognizing knee abnormalities compared to spectrogram and persistence spectrum techniques. The CNN model used in the study was effective in detecting knee abnormalities when fed with 2D projected images of sEMG signals, achieving an accuracy of 86.4%.

Discussion

The superior performance of the scalogram technique can be attributed to its multi-scale analysis capabilities and the avoidance of window size selection required by the STFT method. The CNN model improved classification of lower limb muscle abnormality by utilizing the mapped 2D image format of sEMG frequency-time instead of time segments of sEMG. The technique of mapping sEMG signals to a 2D space and analyzing them with a CNN model could potentially be used as a clinical tool for detecting muscle abnormalities during various types of movement.

Conclusion

Scalogram image representation, using continuous wavelet transform (CWT), provides significantly better performance in recognizing knee abnormalities compared to spectrogram and persistence spectrum. The CNN model used for detecting knee abnormalities showed improved classification performance when processing 2D images rather than 1D time-domain signals. This technique could be a potential clinical tool for detecting muscle abnormalities during various types of movement.

Prospects

  • The technique could be developed into a clinical tool for detecting muscle abnormalities during various types of movement.
  • Further investigation into a more diverse range of muscles could enhance diagnosis accuracy.
  • The paper proposes using automatic deep feature extraction in learning models without additional feature extraction for future work.

Critical Evaluation

Strength

  • The study conducts a comparative experiment between three time-frequency representation techniques (scalogram, spectrogram, and persistence spectrum) for analyzing surface electromyography (sEMG) signals related to lower limb muscle activity.
  • Utilizes a convolutional neural network (CNN) model to classify knee abnormalities from sEMG signals, demonstrating the effectiveness of deep learning in medical diagnostics.
  • Employs data augmentation techniques to increase the dataset size, enhancing the robustness and accuracy of the CNN model.
  • Finds that the scalogram image representation outperforms the other two techniques in recognizing knee abnormalities, highlighting the potential of scalogram for future diagnostic applications.

Weakness

  • sEMG signals are complex and can be easily disturbed by external artifacts like muscle motion or electrode placement.
  • The study focuses on a specific set of muscles and movements, which may limit the generalization of the findings to other muscles and types of movement.
  • While the study proposes a potential clinical tool for detecting muscle abnormalities, further research is needed to validate its effectiveness in clinical settings.

Biases and Issues

  • Although not explicitly stated in the paper, factors such as the size and diversity of the dataset, the choice of muscles analyzed, and the robustness of the CNN model against overfitting and noise should be considered.

Personal Notes

  • Traditional methods of feature extraction may not be sufficient because:
    • sEMG signals lack discriminative features.
    • Signal complexity.
    • Susceptibility to external artifacts or noises like muscle motion or electrode placement.
  • Automatic deep feature extraction in training the model without the need for additional feature extraction processes.
  • Converting 1-D sEMG signals into a 2-D time-frequency space may enhance the detection of muscle activity caused by nerve compression resulting from herniated/bulging discs.